分析｜AI 音视频双轨解析

Most vision nodes describe a video. TikpanGeminiVideoAnalystNode is built to decompose one: it takes video frames and an optional audio track, and produces a professional 分镜 (shot-by-shot) report - 机位 (camera position), 动作, 光影, 场景 - pitched at "好莱坞级" (Hollywood-grade) in the author's own description. The pack's README frames it as the front half of a two-node pipeline: analyze here, then feed the report into the Grok reconstruction node to rewrite the video with matching prompts. If you're reverse-engineering a competitor's ad or turning a reference video into a shootable brief, this is the node.

How it works

Notably, it does not take a video file. The required input is 视频流_IMAGE - a frame sequence from a LoadVideo-style node - plus the source 视频帧率FPS (default 24) so the node can convert frame count back into timestamps. Internally it does automatic frame extraction + an "8MB 智能熔断" (smart chunking cap on the media payload) + synchronized audio analysis when you connect the optional 音频流_AUDIO track. The analysis model is a dropdown - gemini-3.1-flash (fast, default), gemini-3.1-pro (detailed), gpt-5.4-mini, or gpt-4o (stronger at text) - all through the Tikpan relay. The single output is 📄_专业分镜与脚本报告 (STRING), which wires straight into a text viewer or the Grok reconstruct node.

The inputs that actually matter

• API_密钥 - sk- from tikpan.com.
• 视频流_IMAGE - the required frames. This is a real ComfyUI-native design: extract frames locally (free, no upload of a whole file), analyze only what's needed. If you're used to other nodes swallowing a video path, unlearn that here.
• 视频帧率FPS - tell it the source frame rate so the report's timestamps are honest. 24 is the default; an 8 fps source fed in as 24 gives you nonsense timecodes.
• 分析模型 - the quality/cost tradeoff. Flash for drafts, pro for the keepers.
• 音频流_AUDIO - optional, but it's the "双轨" in the name: connecting the audio lets the model follow dialogue and music alongside the visuals. The two-track sync is the feature.
• 重点分析要求 - tell it what you care about (运镜/光影/情绪/带货卖点), default focuses on physical rules, lighting, and micro-expressions.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/htrert/ComfyUI-Tikpan-Pro

restart (or Manager → "Tikpan"). No models to download; the heavy lifting is the relay call plus local frame handling.

Where people get burned

• Wrong input type. Feeding a video file path instead of an IMAGE frame batch fails at the socket. Load frames with a video-frame node first. The ~18MB/48MB inline-video conveniences of other nodes don't exist here - this one is frames-only by design.
• FPS mismatches. Wrong 视频帧率FPS → wrong timestamps in the report. Set it to the actual source.
• Skipping the audio track. The single-track report misses the dialogue/voiceover layer - you're paying for dual-track, so connect the 音频流_AUDIO if the video has sound.
• Model choice = cost. Pro on a long clip is the expensive combo; flash is fine for a first pass.

The honest pitch: it's an analysis node that respects ComfyUI's local-frame workflow and hands you a report you can actually act on - and if you pair it with the Grok reconstruct node, it becomes the "clone this video's grammar" pipeline the README promises.